Repository navigation
v0.5.0
Open Science v0.5.0
This release is about the compute and the agent underneath your workbench — making both pluggable, reproducible, and yours. The notebook layer grows from a single Python kernel into a real multi-kernel runtime: warm Python, R, and REPL control-plane kernels, app-managed conda environments with offline provisioning and bring-your-own interpreters, and crash-recoverable run history. The agent gains a third framework backend, Codex, usable beyond OpenAI through a loopback provider gateway. Providers add StepFun, per-model multimodal image input, and clearer protocol-named labels. And every file preview now opens full-screen. All self-hosted, on your own machine.
Open Science is an open-source, model-agnostic AI workbench for scientific discovery — a self-hosted desktop app pairing a plan-and-execute agent with a persistent, sandboxed compute kernel and durable project/session storage.
This is a compute-and-backend release: it turns the notebook kernel into a reproducible multi-kernel runtime with managed environments, adds a third agent-framework backend, broadens provider and image-input support, and hardens the artifact, reviewer, and attachment paths. A fully model-agnostic gateway, artifact provenance, and the public skills commons remain ahead on the Roadmap.
✨ Highlights
- A real multi-kernel notebook runtime. The one-shot Python executor is replaced by a persistent runtime registry: warm Python, R, and REPL control-plane kernels routed through a single per-session binding, with cross-kernel handoff through a shared workspace channel (the control-plane REPL fetches connector data and hands files to the Python/R kernels), app-managed conda environments provisioned from a version-pinned, SHA256-verified offline bundle (with a configurable mirror), bring-your-own interpreter discovery, a disable/drain lifecycle with no silent fallback, and a crash-recoverable operation journal that reconciles kernels and environments on startup. R is managed-only in this first version. (#223)
- A third agent backend: Codex. Pick Codex in onboarding or Settings alongside Claude Code and OpenCode. It drives providers that speak the OpenAI Responses API natively, and reaches Chat Completions providers (DeepSeek, Kimi, Xiaomi MIMO, OpenRouter, GLM/Z.AI, custom gateways) through a loopback, path-aware translation gateway. The provider key travels over the ACP handshake into an ephemeral in-memory store — never written to config, the on-disk
config.toml, or the process environment. (#212) - StepFun, multimodal image input, and clearer provider labels. Add StepFun as an official vendor (Global + China regions, Messages / Chat Completions / Responses endpoints,
step-3.7-flashmultimodal). Models now carry per-model image-input capability metadata, so a vision-capable model no longer wrongly refuses a pasted screenshot. API-format labels read by protocol now — Messages API, Chat Completions API, Responses API — and the Anthropic vendor is relabeled from Claude to Anthropic. (#230, #222) - Full-screen file previews. Open any preview at 90% of the viewport, from both the right-side preview panel and the Files tab, with a header consistent with the compact view. (#229)
🚀 New Features
- Notebook runtime registry — persistent Python/R/REPL kernels routed per session binding, with cross-kernel handoff through a shared workspace channel; app-managed conda environments with version-pinned, SHA256-verified offline provisioning, a configurable package mirror, named-environment create/remove, and bring-your-own interpreter discovery; a disable-drains-then-revokes lifecycle with no silent fallback; and a crash-recoverable operation journal with startup reconciliation and child-PID supervision. Includes a Settings › Runtimes panel and onboarding environment setup. (#223)
- Codex agent-framework backend — select, install, and detect Codex from within the app, driven over ACP; native OpenAI Responses providers run direct, Chat-Completions providers run through a loopback Responses↔Chat-Completions gateway that coalesces parallel tool calls, round-trips reasoning content, and exposes MCP connectors as namespaced tool aliases. Credentials are isolated in an ephemeral in-memory store and encrypted at rest via the OS credential vault. (#212)
- StepFun provider + protocol-named endpoint labels — StepFun as an official vendor across three endpoints and two regions, with the provider picker reordered (DeepSeek → GLM → Kimi → MiniMax → StepFun → Xiaomi, OpenRouter last), API-format labels renamed to protocol names, and the Anthropic vendor rebranded. (#230)
- Per-model multimodal capability metadata — the official provider registry now records which models accept image input, and image-support detection queries that registry, so multimodal models stop rejecting attached or pasted images. (#222)
- Full-screen file previews — a 90vw × 90vh preview dialog from the preview panel and the Files tab, preserving GENERATED-message preview behavior. (#229)
🔧 Improvements
- Artifact and reviewer lifecycle hardening: safer artifact handling across parallel sessions, orphan cleanup, crash recovery, and reviewer-trust boundaries. (#221)
- Large attachments are no longer inlined whole into the prompt context — oversized text is capped with a preview plus a resource link, keeping requests under provider limits. (#218)
- Inline image payloads are held under the provider's 5 MB base64 per-image cap, so image attachments don't overflow the request. (#225)
- Image attachments are delivered to OpenCode as pixels rather than being dropped, fixing image input on that backend. (#224)
- Skill imports from GitHub route through Electron's proxy-aware
net.fetch, so imports succeed behind a system proxy where Node'sfetchpreviously returned 403. (#215) - Cancelling an update now aborts the in-flight download instead of letting it run to completion. (#216)
🐛 Bug Fixes
- Switching the data-root pointer during onboarding now creates the target folder first, fixing a false "Data folder not found" after onboarding to a custom empty location. (#226)
- Markdown table columns are sized to their content instead of stretching, so rendered tables read correctly. (#220)
open-science startno longer hangs for ~30 seconds when the app is already running — the second instance is routed to the running app. (#214)
📦 Install
Requirements: macOS 12+ (Apple Silicon or Intel), Linux x64, or Windows 10/11 x64. On first run the app's onboarding wizard checks your environment and can install and configure a Claude-compatible model backend for you. Once installed, the app updates itself in place.
Download the build for your platform from the Assets below:
| Platform | File |
|---|---|
| macOS (Apple Silicon) | open-science-0.5.0-mac-arm64.dmg |
| macOS (Intel) | open-science-0.5.0-mac-x64.dmg |
| Linux | open-science-0.5.0-linux-x86_64.AppImage / .deb |
| Windows | open-science-0.5.0-win-x64-setup.exe |
macOS — first launch. Release builds are Developer ID signed and notarized by Apple, so they open like any other app. (A .app you build yourself is not notarized — see README → Building From Source.)
Windows — first launch (unsigned build). No Authenticode certificate yet, so SmartScreen shows a bypassable "unrecognized app" prompt (More info → Run anyway).
Build from source instead:
npm install
npm run build:mac # or: build:linux / build:win
🧭 What's in this release (maturity)
Aligned with the PRD and Roadmap:
- ✅ Implemented: desktop shell (Electron + React + TS) with a tray/single-instance lifecycle, reachable as a desktop app, an optional localhost web UI, a headless backend mode, and an
open-scienceCLI daemon; ACP agent runtime on a pluggable framework backend (Claude Code default, with OpenCode and Codex as alternatives) whose app-managed runtimes are installable, switchable, and uninstallable from Settings; an opt-in reviewer with a bounded fix loop; a persistent multi-kernel notebook runtime — warm Python/R/REPL kernels routed per session binding, app-managed conda environments with version-pinned offline provisioning and bring-your-own interpreters, and crash-recoverable run history; project + session persistence; artifacts MCP server; file renderers + notebook/session-notebook viewers + a full-screen preview mode + offline molecule viewer; multi-provider config (built-in vendors Anthropic (Claude), DeepSeek, Zhipu AI (GLM), MiniMax, Kimi (Moonshot), Kimi For Coding, StepFun, Xiaomi MIMO, plus the OpenRouter aggregation gateway, custom gateways, and local Claude) with per-model image-input metadata, over Anthropic-compatible endpoints on any backend and OpenAI-compatible / Responses endpoints via OpenCode and Codex; signed-and-notarized macOS packaging plus Windows/Linux, with in-place auto-update and a configurable data-storage location; file-based agent skills with a fuzzy/selector and resilient batch import; 24 built-in connectors (23 upstream-aligned life-science connectors at 200+ tools + the molecule viewer) plus custom MCP servers; a guided first-run onboarding wizard; and per-conversation approval profiles. - 🚧 Partial: the new multi-kernel runtime ships R managed-only in this first version, and environment snapshotting/export is not built yet; the CLI/browser surface is a first entry point, not yet a scriptable/embeddable SDK; provider choice is still gated on the selected framework's endpoint compatibility (Claude Code speaks Anthropic; OpenCode and Codex add OpenAI-compatible / Responses) rather than one unified gateway, and Codex's per-vendor live function-tool verification is an ongoing release gate (DeepSeek verified live so far); skills are local and file-based (no shared/public commons, cross-machine forking, or version pinning yet); the reviewer verifies a turn's own record but does not yet apply domain-specific citation/units/statistics rubrics.
- 🗺️ Roadmap (not in this build): public skills commons and savable specialist roles (rest of Phase 3); a fully model-agnostic gateway with unified per-agent routing and a full CLI/SDK entry point (rest of Phase 1); artifact versioning + provenance chain, environment snapshotting/export, and specialist sub-agents (rest of Phase 2); remote compute and the rest of the security stack — network sandbox, credential vault, scoped permission tiers (Phase 4).
🐢 Known Limitations
- R is managed-only in this first runtime version. You get a warm R kernel against an app-managed environment; a bring-your-own R interpreter path is not built yet.
- No environment snapshotting or export yet. You can create, bind, and remove environments, but capturing a reproducible snapshot/lockfile of one is Phase 2 work.
- Codex is wired ahead of full per-vendor live verification. Codex runs against API-key providers today; per-vendor live function-tool verification is an ongoing release gate (DeepSeek is verified live so far), and reusing an existing ChatGPT / local Codex account login is intentionally out of scope in this build.
- Provider choice is per backend, not one unified gateway. Claude Code reaches Anthropic-compatible endpoints; OpenCode and Codex add OpenAI-compatible / Responses ones. The fully model-agnostic gateway with unified per-agent routing is still open Phase 1 work.
- The web UI is loopback-only, and the CLI is an entry point, not an SDK. Both bind to
127.0.0.1; there is no LAN/remote-access mode, and a scriptable/embeddable programmatic interface is still open Phase 1 work. - Switching backends can't carry a live session over. Open conversations keep their messages and get replayed to the new backend, but in-flight tool state is not migrated — switching starts a fresh agent session.
- Skills are local only. You can create, edit, and import them on one machine, but there is no shared/public commons, no cross-machine forking, and no user-facing version pinning yet.
- The reviewer is opt-in and record-scoped. It traces a turn's own claims against its transcript/logs/artifacts; it is off by default and does not yet run domain-specific citation/units/statistics checks.
- Windows builds are unsigned. Windows has no Authenticode cert yet, so it shows a bypassable SmartScreen prompt on first launch. (macOS is notarized.)
- No provenance chain / artifact versioning yet. Lands in Phase 2.
- No local GPU compute backend. The bundled biomodel skills (AlphaFold2, ESMFold2, DiffDock, …) ship as reference/instruction docs; this build can't execute their GPU/model tooling locally.
- No multi-user / real-time collaboration. Sharing is via export/import.
🙏 Acknowledgements
Thanks to @ewen-poch, @wen2zhou, and @roxi3906 (first contribution — full-screen previews and markdown table sizing), and to everyone in Discord, X, and Discussions.
What's Changed
- fix(cli): route
open-science startto the running app instead of hanging by @ewen-poch in #214 - fix(skills): route GitHub import through proxy-aware net.fetch by @ewen-poch in #215
- feat(agent): add Codex ACP support with provider gateway by @ewen-poch in #212
- fix(acp): avoid inlining large attachments into prompt context by @ewen-poch in #218
- feat(providers): add multimodal image support metadata to model registry by @ewen-poch in #222
- fix(update): abort the download when the user cancels (#216) by @ewen-poch in #219
- fix(renderer): size markdown table columns to content by @roxi3906 in #220
- fix(acp): deliver image attachments as pixels to opencode by @ewen-poch in #224
- fix(uploads): keep inline image payload under the 5MB base64 provider limit by @ewen-poch in #225
- fix(storage): create data root on onboarding pointer switch by @ewen-poch in #226
- fix(artifacts): harden the artifact lifecycle (parallel sessions, orphans, crash recovery, reviewer trust) by @ewen-poch in #221
- feat(providers): add StepFun and rename endpoint/vendor labels by @ewen-poch in #230
- feat: add full-screen file previews by @roxi3906 in #229
- feat(notebook): v4 runtime registry — persistent kernels, bindings, offline provisioning, crash recovery by @ewen-poch in #223
- chore: bump version to 0.5.0 by @ewen-poch in #233
Full Changelog: v0.4.1...v0.5.0